Towards Practical Autoconstructive Evolution: Self-Evolution of Problem-Solving Genetic Programming Systems

Abstract

Most genetic programming systems use hard-codedgenetic operators that are applied according touser-specified parameters. Because it is unlikely thatthe provided operators or the default parameters willbe ideal for all problems or all programrepresentations, practitioners often devoteconsiderable energy to experimentation withalternatives. Attempts to bring choices about operatorsand parameters under evolutionary control, throughself-adaptative algorithms or meta-genetic programming,have been explored in the literature and have producedinteresting results. However, no systems based on suchprinciples have yet been demonstrated to have greaterpractical problem-solving power than the more-standardalternatives. This chapter explores the prospects forextending the practical power of genetic programmingthrough the refinement of an approach calledautoconstructive evolution, in which the algorithmsused for the reproduction and variation of evolvingprograms are encoded in the programs themselves, andare thereby subject to variation and evolution intandem with their problem-solving components. Wepresent the motivation for the autoconstructiveevolution approach, show how it can be instantiatedusing the Push programming language, summarise previousresults with the Pushpop system, outline the morerecent AutoPush system, and chart a course for futurework focused on the production of practical systemsthat can solve hard problems.